An AI strategy without a named owner and a P&L line is theater, not strategy. I have reviewed AI roadmaps for Fortune-500 companies and growth-stage startups alike, and the pattern is depressingly consistent: a polished deck, an impressive list of use cases, and precisely zero accountability for whether any of it happens.
The deck gets presented to the board. The board nods approvingly. Everyone leaves feeling like they are an AI company now. Six months later, the same deck gets refreshed with new logos from different vendors and presented again.
This is not a technology failure. It is a governance failure dressed up as a technology strategy.
What a Real AI Strategy Looks Like
A real AI strategy has four components. Most companies have one, maybe two.
1. A named executive owner — not "the AI Committee", not "IT and the CDO jointly". One person who stands up in the monthly leadership review and says: this is on me.
2. A P&L line — investment, expected return, measurement cadence. If you cannot tell me what success looks like in numbers, you do not have a strategy. You have an aspiration.
3. A portfolio of bets, staged by maturity — quick wins shipping in 90 days, medium-term platform investments, and longer-horizon bets that may not pay off for 18 months. All three horizons simultaneously, not sequentially.
4. A production discipline — the explicit process by which an experiment becomes a production system. Most organisations have none. Pilots succeed and then die because nobody owns the engineering effort to operationalise them.
Why AI Strategies Fail: The Honest Diagnosis
I spend a lot of time in post-mortems for stalled AI programmes. The failures cluster around three root causes.
Diffuse ownership disguised as collaboration
"Everyone owns AI" means nobody owns AI. The CDO owns the data strategy. The CTO owns the platform. The business units own the use cases. The AI Centre of Excellence owns... advisory? The result is a game of hot potato where every decision requires a meeting of all four parties and nothing moves fast enough to survive contact with reality.
The companies making real progress have drawn a hard line: one person owns the AI P&L and has authority to make resource allocation decisions without a committee vote.
Pilots as the end state, not the beginning
I have seen companies run 40 pilots and ship zero production systems. The pilot is a safe zone — contained, impressive at demo day, easy to kill without embarrassment. Production is the danger zone: it requires engineering capacity, monitoring, incident response, data pipelines, regulatory sign-off, and ongoing maintenance.
The moment a pilot succeeds, the organisation should be asking: "What is the engineering cost to make this production-grade?" Most organisations ask instead: "What is the next pilot we can run?"
Vendor-driven roadmaps disguised as strategy
When a strategy reads like the feature list of the last vendor who got a meeting, it is not a strategy. It is a procurement wish list. Vendors are excellent at identifying opportunities that require their products. They are not optimising for your business outcomes — they are optimising for their contract value.
A proper AI strategy is vendor-agnostic at the level of business outcomes and deliberately selective at the level of implementation choices. The question is not "which AI platform should we adopt?" The question is "which business problems, if solved with AI, would create the most durable competitive advantage?"
The pilot trap is seductive for a reason
Pilots generate visible activity, create good conference talks, and keep vendors happy. They rarely generate consequences if they fail. This is exactly why organisations default to them. Strategy requires making hard choices about where to concentrate resources — and being accountable when those choices don't pay off. Pilots let you look strategic without accepting accountability.
What the Companies Compounding Value Are Doing Differently
I work with organisations where AI is genuinely compounding — where this year's investment in data infrastructure makes next year's model deployment faster and cheaper. Here is what they are doing that the others are not.
They treat AI as a product discipline, not a project discipline
Projects have start dates and end dates. Products have product managers, roadmaps, and users. When you treat an AI system as a product, you automatically inherit the disciplines that make software organisations effective: user research, metrics, iteration cycles, and ownership.
The companies still stuck in the slide-deck phase are running AI as a series of projects. Each project has a sponsor, a budget, and a delivery date. When the delivery date passes, the sponsor moves on. The "delivered" system slowly degrades without a product owner to maintain it.
They have a narrow, explicit use-case portfolio
More is not better. The organisations wasting the most money on AI are trying to address fifteen use cases simultaneously with a team that could adequately staff three. The constraint is not ideas — it is engineering capacity and change management bandwidth.
The best AI strategies I have seen are almost boring in their specificity. One company's entire AI strategy for 2026 was: automate the three most expensive manual steps in their underwriting workflow, measure cost per policy, and use the savings to fund the next initiative. Not glamorous. Extremely effective.
They build the data foundation before the models
This sounds obvious. It is apparently not, because the number of organisations trying to build AI capabilities on top of unstructured, inconsistent, ungoverned data is staggering.
A model is only as good as the data it is trained on and the data it receives at inference time. If your CRM has three different spellings of the same customer, if your transactional data lives in six different schemas across four different systems, if your historical records have gaps from migration projects in 2019, no amount of model sophistication will compensate.
The Governance Structure That Actually Works
| Role | Accountability | Decision Authority |
|---|---|---|
| AI Executive Sponsor | Business outcomes, P&L | Investment prioritisation, programme start/stop |
| Head of AI / ML | Technical delivery, platform | Architecture, tooling, team resourcing |
| Business Unit AI Lead | Use-case definition, adoption | Requirements, change management within BU |
| Data Governance Lead | Data quality, compliance | Data access, retention, regulatory sign-off |
| MLOps / Platform Team | Production reliability | Deployment, monitoring, incident response |
The critical column is decision authority. Every row should have a person, not a committee, who can make the listed decisions without requiring a meeting of all five roles. The moment a decision requires all five people to agree, you have bureaucracy, not governance.
The ownership test
Ask this question in your next AI strategy review: "If this initiative fails to deliver in 12 months, who specifically is accountable?" If there is any ambiguity in the answer, you do not have an owner. You have a sponsor. Sponsors can redirect blame. Owners cannot.
The Sequencing That Changes Everything
Most boards want to see AI across the whole business simultaneously. This is the wrong instinct. The organisations that compound value from AI follow a sequencing logic that feels counterintuitive until you have seen it work.
Phase 1 (0–6 months): Pick one business problem with measurable economic value, a willing business owner, and decent data. Ship a production system. Not a pilot — a production system with monitoring, fallback logic, and a support model. Demonstrate that your organisation can execute end-to-end.
Phase 2 (6–18 months): Use what you learned about data quality, deployment, and change management to build shared infrastructure. Feature stores, MLOps pipelines, data quality tooling. The second use case should cost 40% less to ship than the first because you are reusing infrastructure.
Phase 3 (18 months onward): Now you have a platform, a playbook, and evidence of ROI. This is when you expand the portfolio, because now you can actually deliver what you commit to.
The flow looks like this:
The seductive alternative is to announce a sweeping AI transformation programme covering all business units from day one. This generates a lot of activity, fills the calendar with steering committee meetings, and typically produces nothing of lasting value.
The Questions Worth Asking Right Now
Before your next AI strategy review, answer these:
- Who is accountable for AI ROI — by name, not by committee?
- How many of your current "AI initiatives" are in production, serving real users, generating measurable value?
- What is the explicit process for moving a pilot to production? Write it out. If it does not exist, there is your problem.
- What percentage of your AI budget is going to data infrastructure versus model experimentation?
- If your top three AI vendors disappeared tomorrow, would your AI strategy still make sense?
If questions 1 through 3 produce uncomfortable silences, the slide deck is your strategy. The good news is that this is a governance and accountability problem, not a technology problem. It is faster to fix than most organisations assume.
Building an AI capability that actually compounds — with the right ownership structure, platform foundations, and production discipline — is the work I do with technology and executive teams. If your AI programme is stuck in pilot purgatory or you want to build it right from the start, let's talk — book a 30-minute discovery call and we can diagnose where the accountability gaps are.